CFLAliasAnalysis.cpp revision 277323
1212795Sdim//===- CFLAliasAnalysis.cpp - CFL-Based Alias Analysis Implementation ------==// 2212795Sdim// 3212795Sdim// The LLVM Compiler Infrastructure 4212795Sdim// 5212795Sdim// This file is distributed under the University of Illinois Open Source 6212795Sdim// License. See LICENSE.TXT for details. 7212795Sdim// 8212795Sdim//===----------------------------------------------------------------------===// 9212795Sdim// 10212795Sdim// This file implements a CFL-based context-insensitive alias analysis 11212795Sdim// algorithm. It does not depend on types. The algorithm is a mixture of the one 12212795Sdim// described in "Demand-driven alias analysis for C" by Xin Zheng and Radu 13212795Sdim// Rugina, and "Fast algorithms for Dyck-CFL-reachability with applications to 14212795Sdim// Alias Analysis" by Zhang Q, Lyu M R, Yuan H, and Su Z. -- to summarize the 15212795Sdim// papers, we build a graph of the uses of a variable, where each node is a 16212795Sdim// memory location, and each edge is an action that happened on that memory 17212795Sdim// location. The "actions" can be one of Dereference, Reference, Assign, or 18239462Sdim// Assign. 19212795Sdim// 20212795Sdim// Two variables are considered as aliasing iff you can reach one value's node 21212795Sdim// from the other value's node and the language formed by concatenating all of 22212795Sdim// the edge labels (actions) conforms to a context-free grammar. 23212795Sdim// 24212795Sdim// Because this algorithm requires a graph search on each query, we execute the 25212795Sdim// algorithm outlined in "Fast algorithms..." (mentioned above) 26234353Sdim// in order to transform the graph into sets of variables that may alias in 27212795Sdim// ~nlogn time (n = number of variables.), which makes queries take constant 28212795Sdim// time. 29212795Sdim//===----------------------------------------------------------------------===// 30212795Sdim 31212795Sdim#include "StratifiedSets.h" 32212795Sdim#include "llvm/ADT/BitVector.h" 33212795Sdim#include "llvm/ADT/DenseMap.h" 34212795Sdim#include "llvm/ADT/None.h" 35212795Sdim#include "llvm/ADT/Optional.h" 36212795Sdim#include "llvm/Analysis/AliasAnalysis.h" 37212795Sdim#include "llvm/Analysis/Passes.h" 38212795Sdim#include "llvm/IR/Constants.h" 39212795Sdim#include "llvm/IR/Function.h" 40212795Sdim#include "llvm/IR/InstVisitor.h" 41212795Sdim#include "llvm/IR/Instructions.h" 42212795Sdim#include "llvm/IR/ValueHandle.h" 43212795Sdim#include "llvm/Pass.h" 44212795Sdim#include "llvm/Support/Allocator.h" 45212795Sdim#include "llvm/Support/Compiler.h" 46212795Sdim#include "llvm/Support/ErrorHandling.h" 47212795Sdim#include <algorithm> 48212795Sdim#include <cassert> 49212795Sdim#include <forward_list> 50212795Sdim#include <tuple> 51212795Sdim 52212795Sdimusing namespace llvm; 53212795Sdim 54212795Sdim// Try to go from a Value* to a Function*. Never returns nullptr. 55212795Sdimstatic Optional<Function *> parentFunctionOfValue(Value *); 56212795Sdim 57212795Sdim// Returns possible functions called by the Inst* into the given 58212795Sdim// SmallVectorImpl. Returns true if targets found, false otherwise. 59212795Sdim// This is templated because InvokeInst/CallInst give us the same 60212795Sdim// set of functions that we care about, and I don't like repeating 61212795Sdim// myself. 62212795Sdimtemplate <typename Inst> 63212795Sdimstatic bool getPossibleTargets(Inst *, SmallVectorImpl<Function *> &); 64212795Sdim 65212795Sdim// Some instructions need to have their users tracked. Instructions like 66212795Sdim// `add` require you to get the users of the Instruction* itself, other 67212795Sdim// instructions like `store` require you to get the users of the first 68212795Sdim// operand. This function gets the "proper" value to track for each 69234353Sdim// type of instruction we support. 70234353Sdimstatic Optional<Value *> getTargetValue(Instruction *); 71212795Sdim 72212795Sdim// There are certain instructions (i.e. FenceInst, etc.) that we ignore. 73212795Sdim// This notes that we should ignore those. 74212795Sdimstatic bool hasUsefulEdges(Instruction *); 75212795Sdim 76212795Sdimconst StratifiedIndex StratifiedLink::SetSentinel = 77212795Sdim std::numeric_limits<StratifiedIndex>::max(); 78212795Sdim 79212795Sdimnamespace { 80212795Sdim// StratifiedInfo Attribute things. 81212795Sdimtypedef unsigned StratifiedAttr; 82212795SdimLLVM_CONSTEXPR unsigned MaxStratifiedAttrIndex = NumStratifiedAttrs; 83212795SdimLLVM_CONSTEXPR unsigned AttrAllIndex = 0; 84212795SdimLLVM_CONSTEXPR unsigned AttrGlobalIndex = 1; 85212795SdimLLVM_CONSTEXPR unsigned AttrFirstArgIndex = 2; 86212795SdimLLVM_CONSTEXPR unsigned AttrLastArgIndex = MaxStratifiedAttrIndex; 87212795SdimLLVM_CONSTEXPR unsigned AttrMaxNumArgs = AttrLastArgIndex - AttrFirstArgIndex; 88212795Sdim 89212795SdimLLVM_CONSTEXPR StratifiedAttr AttrNone = 0; 90212795SdimLLVM_CONSTEXPR StratifiedAttr AttrAll = ~AttrNone; 91212795Sdim 92212795Sdim// \brief StratifiedSets call for knowledge of "direction", so this is how we 93212795Sdim// represent that locally. 94212795Sdimenum class Level { Same, Above, Below }; 95212795Sdim 96212795Sdim// \brief Edges can be one of four "weights" -- each weight must have an inverse 97212795Sdim// weight (Assign has Assign; Reference has Dereference). 98212795Sdimenum class EdgeType { 99212795Sdim // The weight assigned when assigning from or to a value. For example, in: 100212795Sdim // %b = getelementptr %a, 0 101212795Sdim // ...The relationships are %b assign %a, and %a assign %b. This used to be 102212795Sdim // two edges, but having a distinction bought us nothing. 103212795Sdim Assign, 104212795Sdim 105212795Sdim // The edge used when we have an edge going from some handle to a Value. 106212795Sdim // Examples of this include: 107212795Sdim // %b = load %a (%b Dereference %a) 108212795Sdim // %b = extractelement %a, 0 (%a Dereference %b) 109212795Sdim Dereference, 110212795Sdim 111212795Sdim // The edge used when our edge goes from a value to a handle that may have 112212795Sdim // contained it at some point. Examples: 113212795Sdim // %b = load %a (%a Reference %b) 114212795Sdim // %b = extractelement %a, 0 (%b Reference %a) 115212795Sdim Reference 116212795Sdim}; 117212795Sdim 118212795Sdim// \brief Encodes the notion of a "use" 119212795Sdimstruct Edge { 120212795Sdim // \brief Which value the edge is coming from 121212795Sdim Value *From; 122212795Sdim 123212795Sdim // \brief Which value the edge is pointing to 124212795Sdim Value *To; 125212795Sdim 126212795Sdim // \brief Edge weight 127212795Sdim EdgeType Weight; 128212795Sdim 129212795Sdim // \brief Whether we aliased any external values along the way that may be 130212795Sdim // invisible to the analysis (i.e. landingpad for exceptions, calls for 131212795Sdim // interprocedural analysis, etc.) 132212795Sdim StratifiedAttrs AdditionalAttrs; 133212795Sdim 134212795Sdim Edge(Value *From, Value *To, EdgeType W, StratifiedAttrs A) 135234353Sdim : From(From), To(To), Weight(W), AdditionalAttrs(A) {} 136212795Sdim}; 137212795Sdim 138212795Sdim// \brief Information we have about a function and would like to keep around 139212795Sdimstruct FunctionInfo { 140212795Sdim StratifiedSets<Value *> Sets; 141212795Sdim // Lots of functions have < 4 returns. Adjust as necessary. 142212795Sdim SmallVector<Value *, 4> ReturnedValues; 143212795Sdim 144212795Sdim FunctionInfo(StratifiedSets<Value *> &&S, 145212795Sdim SmallVector<Value *, 4> &&RV) 146212795Sdim : Sets(std::move(S)), ReturnedValues(std::move(RV)) {} 147212795Sdim}; 148212795Sdim 149212795Sdimstruct CFLAliasAnalysis; 150212795Sdim 151212795Sdimstruct FunctionHandle : public CallbackVH { 152212795Sdim FunctionHandle(Function *Fn, CFLAliasAnalysis *CFLAA) 153212795Sdim : CallbackVH(Fn), CFLAA(CFLAA) { 154212795Sdim assert(Fn != nullptr); 155212795Sdim assert(CFLAA != nullptr); 156212795Sdim } 157212795Sdim 158212795Sdim virtual ~FunctionHandle() {} 159212795Sdim 160212795Sdim void deleted() override { removeSelfFromCache(); } 161212795Sdim void allUsesReplacedWith(Value *) override { removeSelfFromCache(); } 162212795Sdim 163212795Sdimprivate: 164212795Sdim CFLAliasAnalysis *CFLAA; 165212795Sdim 166212795Sdim void removeSelfFromCache(); 167212795Sdim}; 168212795Sdim 169212795Sdimstruct CFLAliasAnalysis : public ImmutablePass, public AliasAnalysis { 170212795Sdimprivate: 171212795Sdim /// \brief Cached mapping of Functions to their StratifiedSets. 172212795Sdim /// If a function's sets are currently being built, it is marked 173212795Sdim /// in the cache as an Optional without a value. This way, if we 174212795Sdim /// have any kind of recursion, it is discernable from a function 175212795Sdim /// that simply has empty sets. 176212795Sdim DenseMap<Function *, Optional<FunctionInfo>> Cache; 177212795Sdim std::forward_list<FunctionHandle> Handles; 178234353Sdim 179212795Sdimpublic: 180212795Sdim static char ID; 181212795Sdim 182212795Sdim CFLAliasAnalysis() : ImmutablePass(ID) { 183212795Sdim initializeCFLAliasAnalysisPass(*PassRegistry::getPassRegistry()); 184212795Sdim } 185212795Sdim 186212795Sdim virtual ~CFLAliasAnalysis() {} 187212795Sdim 188212795Sdim void getAnalysisUsage(AnalysisUsage &AU) const override { 189212795Sdim AliasAnalysis::getAnalysisUsage(AU); 190212795Sdim } 191212795Sdim 192212795Sdim void *getAdjustedAnalysisPointer(const void *ID) override { 193212795Sdim if (ID == &AliasAnalysis::ID) 194212795Sdim return (AliasAnalysis *)this; 195234353Sdim return this; 196239462Sdim } 197234353Sdim 198234353Sdim /// \brief Inserts the given Function into the cache. 199234353Sdim void scan(Function *Fn); 200239462Sdim 201234353Sdim void evict(Function *Fn) { Cache.erase(Fn); } 202234353Sdim 203234353Sdim /// \brief Ensures that the given function is available in the cache. 204234353Sdim /// Returns the appropriate entry from the cache. 205234353Sdim const Optional<FunctionInfo> &ensureCached(Function *Fn) { 206234353Sdim auto Iter = Cache.find(Fn); 207239462Sdim if (Iter == Cache.end()) { 208239462Sdim scan(Fn); 209234353Sdim Iter = Cache.find(Fn); 210239462Sdim assert(Iter != Cache.end()); 211239462Sdim assert(Iter->second.hasValue()); 212239462Sdim } 213239462Sdim return Iter->second; 214239462Sdim } 215234353Sdim 216234353Sdim AliasResult query(const Location &LocA, const Location &LocB); 217239462Sdim 218234353Sdim AliasResult alias(const Location &LocA, const Location &LocB) override { 219239462Sdim if (LocA.Ptr == LocB.Ptr) { 220234353Sdim if (LocA.Size == LocB.Size) { 221234353Sdim return MustAlias; 222239462Sdim } else { 223234353Sdim return PartialAlias; 224239462Sdim } 225234353Sdim } 226234353Sdim 227239462Sdim // Comparisons between global variables and other constants should be 228239462Sdim // handled by BasicAA. 229239462Sdim if (isa<Constant>(LocA.Ptr) && isa<Constant>(LocB.Ptr)) { 230239462Sdim return MayAlias; 231239462Sdim } 232234353Sdim 233234353Sdim return query(LocA, LocB); 234234353Sdim } 235234353Sdim 236234353Sdim void initializePass() override { InitializeAliasAnalysis(this); } 237234353Sdim}; 238234353Sdim 239234353Sdimvoid FunctionHandle::removeSelfFromCache() { 240239462Sdim assert(CFLAA != nullptr); 241239462Sdim auto *Val = getValPtr(); 242239462Sdim CFLAA->evict(cast<Function>(Val)); 243239462Sdim setValPtr(nullptr); 244239462Sdim} 245239462Sdim 246239462Sdim// \brief Gets the edges our graph should have, based on an Instruction* 247239462Sdimclass GetEdgesVisitor : public InstVisitor<GetEdgesVisitor, void> { 248234353Sdim CFLAliasAnalysis &AA; 249234353Sdim SmallVectorImpl<Edge> &Output; 250239462Sdim 251234353Sdimpublic: 252239462Sdim GetEdgesVisitor(CFLAliasAnalysis &AA, SmallVectorImpl<Edge> &Output) 253234353Sdim : AA(AA), Output(Output) {} 254239462Sdim 255234353Sdim void visitInstruction(Instruction &) { 256234353Sdim llvm_unreachable("Unsupported instruction encountered"); 257239462Sdim } 258234353Sdim 259234353Sdim void visitCastInst(CastInst &Inst) { 260239462Sdim Output.push_back(Edge(&Inst, Inst.getOperand(0), EdgeType::Assign, 261239462Sdim AttrNone)); 262239462Sdim } 263239462Sdim 264239462Sdim void visitBinaryOperator(BinaryOperator &Inst) { 265234353Sdim auto *Op1 = Inst.getOperand(0); 266234353Sdim auto *Op2 = Inst.getOperand(1); 267234353Sdim Output.push_back(Edge(&Inst, Op1, EdgeType::Assign, AttrNone)); 268234353Sdim Output.push_back(Edge(&Inst, Op2, EdgeType::Assign, AttrNone)); 269234353Sdim } 270234353Sdim 271234353Sdim void visitAtomicCmpXchgInst(AtomicCmpXchgInst &Inst) { 272234353Sdim auto *Ptr = Inst.getPointerOperand(); 273234353Sdim auto *Val = Inst.getNewValOperand(); 274234353Sdim Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone)); 275234353Sdim } 276234353Sdim 277239462Sdim void visitAtomicRMWInst(AtomicRMWInst &Inst) { 278239462Sdim auto *Ptr = Inst.getPointerOperand(); 279239462Sdim auto *Val = Inst.getValOperand(); 280239462Sdim Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone)); 281234353Sdim } 282239462Sdim 283234353Sdim void visitPHINode(PHINode &Inst) { 284239462Sdim for (unsigned I = 0, E = Inst.getNumIncomingValues(); I != E; ++I) { 285234353Sdim Value *Val = Inst.getIncomingValue(I); 286234353Sdim Output.push_back(Edge(&Inst, Val, EdgeType::Assign, AttrNone)); 287234353Sdim } 288234353Sdim } 289234353Sdim 290234353Sdim void visitGetElementPtrInst(GetElementPtrInst &Inst) { 291234353Sdim auto *Op = Inst.getPointerOperand(); 292239462Sdim Output.push_back(Edge(&Inst, Op, EdgeType::Assign, AttrNone)); 293239462Sdim for (auto I = Inst.idx_begin(), E = Inst.idx_end(); I != E; ++I) 294234353Sdim Output.push_back(Edge(&Inst, *I, EdgeType::Assign, AttrNone)); 295239462Sdim } 296234353Sdim 297234353Sdim void visitSelectInst(SelectInst &Inst) { 298239462Sdim auto *Condition = Inst.getCondition(); 299234353Sdim Output.push_back(Edge(&Inst, Condition, EdgeType::Assign, AttrNone)); 300239462Sdim auto *TrueVal = Inst.getTrueValue(); 301234353Sdim Output.push_back(Edge(&Inst, TrueVal, EdgeType::Assign, AttrNone)); 302234353Sdim auto *FalseVal = Inst.getFalseValue(); 303234353Sdim Output.push_back(Edge(&Inst, FalseVal, EdgeType::Assign, AttrNone)); 304234353Sdim } 305234353Sdim 306234353Sdim void visitAllocaInst(AllocaInst &) {} 307239462Sdim 308234353Sdim void visitLoadInst(LoadInst &Inst) { 309234353Sdim auto *Ptr = Inst.getPointerOperand(); 310239462Sdim auto *Val = &Inst; 311234353Sdim Output.push_back(Edge(Val, Ptr, EdgeType::Reference, AttrNone)); 312239462Sdim } 313234353Sdim 314234353Sdim void visitStoreInst(StoreInst &Inst) { 315234353Sdim auto *Ptr = Inst.getPointerOperand(); 316234353Sdim auto *Val = Inst.getValueOperand(); 317239462Sdim Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone)); 318234353Sdim } 319239462Sdim 320239462Sdim void visitVAArgInst(VAArgInst &Inst) { 321239462Sdim // We can't fully model va_arg here. For *Ptr = Inst.getOperand(0), it does 322239462Sdim // two things: 323239462Sdim // 1. Loads a value from *((T*)*Ptr). 324239462Sdim // 2. Increments (stores to) *Ptr by some target-specific amount. 325239462Sdim // For now, we'll handle this like a landingpad instruction (by placing the 326239462Sdim // result in its own group, and having that group alias externals). 327239462Sdim auto *Val = &Inst; 328239462Sdim Output.push_back(Edge(Val, Val, EdgeType::Assign, AttrAll)); 329239462Sdim } 330239462Sdim 331239462Sdim static bool isFunctionExternal(Function *Fn) { 332239462Sdim return Fn->isDeclaration() || !Fn->hasLocalLinkage(); 333239462Sdim } 334239462Sdim 335239462Sdim // Gets whether the sets at Index1 above, below, or equal to the sets at 336239462Sdim // Index2. Returns None if they are not in the same set chain. 337239462Sdim static Optional<Level> getIndexRelation(const StratifiedSets<Value *> &Sets, 338239462Sdim StratifiedIndex Index1, 339239462Sdim StratifiedIndex Index2) { 340239462Sdim if (Index1 == Index2) 341239462Sdim return Level::Same; 342239462Sdim 343239462Sdim const auto *Current = &Sets.getLink(Index1); 344239462Sdim while (Current->hasBelow()) { 345239462Sdim if (Current->Below == Index2) 346234353Sdim return Level::Below; 347234353Sdim Current = &Sets.getLink(Current->Below); 348234353Sdim } 349234353Sdim 350239462Sdim Current = &Sets.getLink(Index1); 351234353Sdim while (Current->hasAbove()) { 352234353Sdim if (Current->Above == Index2) 353234353Sdim return Level::Above; 354234353Sdim Current = &Sets.getLink(Current->Above); 355234353Sdim } 356234353Sdim 357234353Sdim return NoneType(); 358239462Sdim } 359239462Sdim 360239462Sdim bool 361239462Sdim tryInterproceduralAnalysis(const SmallVectorImpl<Function *> &Fns, 362234353Sdim Value *FuncValue, 363239462Sdim const iterator_range<User::op_iterator> &Args) { 364239462Sdim const unsigned ExpectedMaxArgs = 8; 365239462Sdim const unsigned MaxSupportedArgs = 50; 366239462Sdim assert(Fns.size() > 0); 367239462Sdim 368234353Sdim // I put this here to give us an upper bound on time taken by IPA. Is it 369234353Sdim // really (realistically) needed? Keep in mind that we do have an n^2 algo. 370234353Sdim if (std::distance(Args.begin(), Args.end()) > (int) MaxSupportedArgs) 371234353Sdim return false; 372234353Sdim 373234353Sdim // Exit early if we'll fail anyway 374234353Sdim for (auto *Fn : Fns) { 375234353Sdim if (isFunctionExternal(Fn) || Fn->isVarArg()) 376234353Sdim return false; 377234353Sdim auto &MaybeInfo = AA.ensureCached(Fn); 378234353Sdim if (!MaybeInfo.hasValue()) 379234353Sdim return false; 380234353Sdim } 381234353Sdim 382234353Sdim SmallVector<Value *, ExpectedMaxArgs> Arguments(Args.begin(), Args.end()); 383234353Sdim SmallVector<StratifiedInfo, ExpectedMaxArgs> Parameters; 384234353Sdim for (auto *Fn : Fns) { 385234353Sdim auto &Info = *AA.ensureCached(Fn); 386234353Sdim auto &Sets = Info.Sets; 387234353Sdim auto &RetVals = Info.ReturnedValues; 388234353Sdim 389234353Sdim Parameters.clear(); 390234353Sdim for (auto &Param : Fn->args()) { 391234353Sdim auto MaybeInfo = Sets.find(&Param); 392234353Sdim // Did a new parameter somehow get added to the function/slip by? 393234353Sdim if (!MaybeInfo.hasValue()) 394234353Sdim return false; 395234353Sdim Parameters.push_back(*MaybeInfo); 396234353Sdim } 397234353Sdim 398234353Sdim // Adding an edge from argument -> return value for each parameter that 399234353Sdim // may alias the return value 400234353Sdim for (unsigned I = 0, E = Parameters.size(); I != E; ++I) { 401234353Sdim auto &ParamInfo = Parameters[I]; 402234353Sdim auto &ArgVal = Arguments[I]; 403234353Sdim bool AddEdge = false; 404234353Sdim StratifiedAttrs Externals; 405234353Sdim for (unsigned X = 0, XE = RetVals.size(); X != XE; ++X) { 406234353Sdim auto MaybeInfo = Sets.find(RetVals[X]); 407234353Sdim if (!MaybeInfo.hasValue()) 408234353Sdim return false; 409234353Sdim 410234353Sdim auto &RetInfo = *MaybeInfo; 411234353Sdim auto RetAttrs = Sets.getLink(RetInfo.Index).Attrs; 412234353Sdim auto ParamAttrs = Sets.getLink(ParamInfo.Index).Attrs; 413234353Sdim auto MaybeRelation = 414234353Sdim getIndexRelation(Sets, ParamInfo.Index, RetInfo.Index); 415234353Sdim if (MaybeRelation.hasValue()) { 416234353Sdim AddEdge = true; 417234353Sdim Externals |= RetAttrs | ParamAttrs; 418234353Sdim } 419234353Sdim } 420234353Sdim if (AddEdge) 421234353Sdim Output.push_back(Edge(FuncValue, ArgVal, EdgeType::Assign, 422234353Sdim StratifiedAttrs().flip())); 423234353Sdim } 424234353Sdim 425234353Sdim if (Parameters.size() != Arguments.size()) 426234353Sdim return false; 427234353Sdim 428234353Sdim // Adding edges between arguments for arguments that may end up aliasing 429239462Sdim // each other. This is necessary for functions such as 430239462Sdim // void foo(int** a, int** b) { *a = *b; } 431234353Sdim // (Technically, the proper sets for this would be those below 432234353Sdim // Arguments[I] and Arguments[X], but our algorithm will produce 433239462Sdim // extremely similar, and equally correct, results either way) 434239462Sdim for (unsigned I = 0, E = Arguments.size(); I != E; ++I) { 435234353Sdim auto &MainVal = Arguments[I]; 436234353Sdim auto &MainInfo = Parameters[I]; 437234353Sdim auto &MainAttrs = Sets.getLink(MainInfo.Index).Attrs; 438234353Sdim for (unsigned X = I + 1; X != E; ++X) { 439234353Sdim auto &SubInfo = Parameters[X]; 440234353Sdim auto &SubVal = Arguments[X]; 441234353Sdim auto &SubAttrs = Sets.getLink(SubInfo.Index).Attrs; 442234353Sdim auto MaybeRelation = 443234353Sdim getIndexRelation(Sets, MainInfo.Index, SubInfo.Index); 444234353Sdim 445234353Sdim if (!MaybeRelation.hasValue()) 446234353Sdim continue; 447234353Sdim 448234353Sdim auto NewAttrs = SubAttrs | MainAttrs; 449234353Sdim Output.push_back(Edge(MainVal, SubVal, EdgeType::Assign, NewAttrs)); 450234353Sdim } 451234353Sdim } 452234353Sdim } 453234353Sdim return true; 454234353Sdim } 455234353Sdim 456234353Sdim template <typename InstT> void visitCallLikeInst(InstT &Inst) { 457234353Sdim SmallVector<Function *, 4> Targets; 458234353Sdim if (getPossibleTargets(&Inst, Targets)) { 459234353Sdim if (tryInterproceduralAnalysis(Targets, &Inst, Inst.arg_operands())) 460234353Sdim return; 461234353Sdim // Cleanup from interprocedural analysis 462234353Sdim Output.clear(); 463212795Sdim } 464212795Sdim 465234353Sdim for (Value *V : Inst.arg_operands()) 466234353Sdim Output.push_back(Edge(&Inst, V, EdgeType::Assign, AttrAll)); 467212795Sdim } 468212795Sdim 469212795Sdim void visitCallInst(CallInst &Inst) { visitCallLikeInst(Inst); } 470212795Sdim 471212795Sdim void visitInvokeInst(InvokeInst &Inst) { visitCallLikeInst(Inst); } 472234353Sdim 473234353Sdim // Because vectors/aggregates are immutable and unaddressable, 474212795Sdim // there's nothing we can do to coax a value out of them, other 475212795Sdim // than calling Extract{Element,Value}. We can effectively treat 476212795Sdim // them as pointers to arbitrary memory locations we can store in 477212795Sdim // and load from. 478212795Sdim void visitExtractElementInst(ExtractElementInst &Inst) { 479212795Sdim auto *Ptr = Inst.getVectorOperand(); 480212795Sdim auto *Val = &Inst; 481212795Sdim Output.push_back(Edge(Val, Ptr, EdgeType::Reference, AttrNone)); 482212795Sdim } 483212795Sdim 484212795Sdim void visitInsertElementInst(InsertElementInst &Inst) { 485212795Sdim auto *Vec = Inst.getOperand(0); 486212795Sdim auto *Val = Inst.getOperand(1); 487212795Sdim Output.push_back(Edge(&Inst, Vec, EdgeType::Assign, AttrNone)); 488212795Sdim Output.push_back(Edge(&Inst, Val, EdgeType::Dereference, AttrNone)); 489212795Sdim } 490212795Sdim 491 void visitLandingPadInst(LandingPadInst &Inst) { 492 // Exceptions come from "nowhere", from our analysis' perspective. 493 // So we place the instruction its own group, noting that said group may 494 // alias externals 495 Output.push_back(Edge(&Inst, &Inst, EdgeType::Assign, AttrAll)); 496 } 497 498 void visitInsertValueInst(InsertValueInst &Inst) { 499 auto *Agg = Inst.getOperand(0); 500 auto *Val = Inst.getOperand(1); 501 Output.push_back(Edge(&Inst, Agg, EdgeType::Assign, AttrNone)); 502 Output.push_back(Edge(&Inst, Val, EdgeType::Dereference, AttrNone)); 503 } 504 505 void visitExtractValueInst(ExtractValueInst &Inst) { 506 auto *Ptr = Inst.getAggregateOperand(); 507 Output.push_back(Edge(&Inst, Ptr, EdgeType::Reference, AttrNone)); 508 } 509 510 void visitShuffleVectorInst(ShuffleVectorInst &Inst) { 511 auto *From1 = Inst.getOperand(0); 512 auto *From2 = Inst.getOperand(1); 513 Output.push_back(Edge(&Inst, From1, EdgeType::Assign, AttrNone)); 514 Output.push_back(Edge(&Inst, From2, EdgeType::Assign, AttrNone)); 515 } 516}; 517 518// For a given instruction, we need to know which Value* to get the 519// users of in order to build our graph. In some cases (i.e. add), 520// we simply need the Instruction*. In other cases (i.e. store), 521// finding the users of the Instruction* is useless; we need to find 522// the users of the first operand. This handles determining which 523// value to follow for us. 524// 525// Note: we *need* to keep this in sync with GetEdgesVisitor. Add 526// something to GetEdgesVisitor, add it here -- remove something from 527// GetEdgesVisitor, remove it here. 528class GetTargetValueVisitor 529 : public InstVisitor<GetTargetValueVisitor, Value *> { 530public: 531 Value *visitInstruction(Instruction &Inst) { return &Inst; } 532 533 Value *visitStoreInst(StoreInst &Inst) { return Inst.getPointerOperand(); } 534 535 Value *visitAtomicCmpXchgInst(AtomicCmpXchgInst &Inst) { 536 return Inst.getPointerOperand(); 537 } 538 539 Value *visitAtomicRMWInst(AtomicRMWInst &Inst) { 540 return Inst.getPointerOperand(); 541 } 542 543 Value *visitInsertElementInst(InsertElementInst &Inst) { 544 return Inst.getOperand(0); 545 } 546 547 Value *visitInsertValueInst(InsertValueInst &Inst) { 548 return Inst.getAggregateOperand(); 549 } 550}; 551 552// Set building requires a weighted bidirectional graph. 553template <typename EdgeTypeT> class WeightedBidirectionalGraph { 554public: 555 typedef std::size_t Node; 556 557private: 558 const static Node StartNode = Node(0); 559 560 struct Edge { 561 EdgeTypeT Weight; 562 Node Other; 563 564 Edge(const EdgeTypeT &W, const Node &N) 565 : Weight(W), Other(N) {} 566 567 bool operator==(const Edge &E) const { 568 return Weight == E.Weight && Other == E.Other; 569 } 570 571 bool operator!=(const Edge &E) const { return !operator==(E); } 572 }; 573 574 struct NodeImpl { 575 std::vector<Edge> Edges; 576 }; 577 578 std::vector<NodeImpl> NodeImpls; 579 580 bool inbounds(Node NodeIndex) const { return NodeIndex < NodeImpls.size(); } 581 582 const NodeImpl &getNode(Node N) const { return NodeImpls[N]; } 583 NodeImpl &getNode(Node N) { return NodeImpls[N]; } 584 585public: 586 // ----- Various Edge iterators for the graph ----- // 587 588 // \brief Iterator for edges. Because this graph is bidirected, we don't 589 // allow modificaiton of the edges using this iterator. Additionally, the 590 // iterator becomes invalid if you add edges to or from the node you're 591 // getting the edges of. 592 struct EdgeIterator : public std::iterator<std::forward_iterator_tag, 593 std::tuple<EdgeTypeT, Node *>> { 594 EdgeIterator(const typename std::vector<Edge>::const_iterator &Iter) 595 : Current(Iter) {} 596 597 EdgeIterator(NodeImpl &Impl) : Current(Impl.begin()) {} 598 599 EdgeIterator &operator++() { 600 ++Current; 601 return *this; 602 } 603 604 EdgeIterator operator++(int) { 605 EdgeIterator Copy(Current); 606 operator++(); 607 return Copy; 608 } 609 610 std::tuple<EdgeTypeT, Node> &operator*() { 611 Store = std::make_tuple(Current->Weight, Current->Other); 612 return Store; 613 } 614 615 bool operator==(const EdgeIterator &Other) const { 616 return Current == Other.Current; 617 } 618 619 bool operator!=(const EdgeIterator &Other) const { 620 return !operator==(Other); 621 } 622 623 private: 624 typename std::vector<Edge>::const_iterator Current; 625 std::tuple<EdgeTypeT, Node> Store; 626 }; 627 628 // Wrapper for EdgeIterator with begin()/end() calls. 629 struct EdgeIterable { 630 EdgeIterable(const std::vector<Edge> &Edges) 631 : BeginIter(Edges.begin()), EndIter(Edges.end()) {} 632 633 EdgeIterator begin() { return EdgeIterator(BeginIter); } 634 635 EdgeIterator end() { return EdgeIterator(EndIter); } 636 637 private: 638 typename std::vector<Edge>::const_iterator BeginIter; 639 typename std::vector<Edge>::const_iterator EndIter; 640 }; 641 642 // ----- Actual graph-related things ----- // 643 644 WeightedBidirectionalGraph() {} 645 646 WeightedBidirectionalGraph(WeightedBidirectionalGraph<EdgeTypeT> &&Other) 647 : NodeImpls(std::move(Other.NodeImpls)) {} 648 649 WeightedBidirectionalGraph<EdgeTypeT> & 650 operator=(WeightedBidirectionalGraph<EdgeTypeT> &&Other) { 651 NodeImpls = std::move(Other.NodeImpls); 652 return *this; 653 } 654 655 Node addNode() { 656 auto Index = NodeImpls.size(); 657 auto NewNode = Node(Index); 658 NodeImpls.push_back(NodeImpl()); 659 return NewNode; 660 } 661 662 void addEdge(Node From, Node To, const EdgeTypeT &Weight, 663 const EdgeTypeT &ReverseWeight) { 664 assert(inbounds(From)); 665 assert(inbounds(To)); 666 auto &FromNode = getNode(From); 667 auto &ToNode = getNode(To); 668 FromNode.Edges.push_back(Edge(Weight, To)); 669 ToNode.Edges.push_back(Edge(ReverseWeight, From)); 670 } 671 672 EdgeIterable edgesFor(const Node &N) const { 673 const auto &Node = getNode(N); 674 return EdgeIterable(Node.Edges); 675 } 676 677 bool empty() const { return NodeImpls.empty(); } 678 std::size_t size() const { return NodeImpls.size(); } 679 680 // \brief Gets an arbitrary node in the graph as a starting point for 681 // traversal. 682 Node getEntryNode() { 683 assert(inbounds(StartNode)); 684 return StartNode; 685 } 686}; 687 688typedef WeightedBidirectionalGraph<std::pair<EdgeType, StratifiedAttrs>> GraphT; 689typedef DenseMap<Value *, GraphT::Node> NodeMapT; 690} 691 692// -- Setting up/registering CFLAA pass -- // 693char CFLAliasAnalysis::ID = 0; 694 695INITIALIZE_AG_PASS(CFLAliasAnalysis, AliasAnalysis, "cfl-aa", 696 "CFL-Based AA implementation", false, true, false) 697 698ImmutablePass *llvm::createCFLAliasAnalysisPass() { 699 return new CFLAliasAnalysis(); 700} 701 702//===----------------------------------------------------------------------===// 703// Function declarations that require types defined in the namespace above 704//===----------------------------------------------------------------------===// 705 706// Given an argument number, returns the appropriate Attr index to set. 707static StratifiedAttr argNumberToAttrIndex(StratifiedAttr); 708 709// Given a Value, potentially return which AttrIndex it maps to. 710static Optional<StratifiedAttr> valueToAttrIndex(Value *Val); 711 712// Gets the inverse of a given EdgeType. 713static EdgeType flipWeight(EdgeType); 714 715// Gets edges of the given Instruction*, writing them to the SmallVector*. 716static void argsToEdges(CFLAliasAnalysis &, Instruction *, 717 SmallVectorImpl<Edge> &); 718 719// Gets the "Level" that one should travel in StratifiedSets 720// given an EdgeType. 721static Level directionOfEdgeType(EdgeType); 722 723// Builds the graph needed for constructing the StratifiedSets for the 724// given function 725static void buildGraphFrom(CFLAliasAnalysis &, Function *, 726 SmallVectorImpl<Value *> &, NodeMapT &, GraphT &); 727 728// Builds the graph + StratifiedSets for a function. 729static FunctionInfo buildSetsFrom(CFLAliasAnalysis &, Function *); 730 731static Optional<Function *> parentFunctionOfValue(Value *Val) { 732 if (auto *Inst = dyn_cast<Instruction>(Val)) { 733 auto *Bb = Inst->getParent(); 734 return Bb->getParent(); 735 } 736 737 if (auto *Arg = dyn_cast<Argument>(Val)) 738 return Arg->getParent(); 739 return NoneType(); 740} 741 742template <typename Inst> 743static bool getPossibleTargets(Inst *Call, 744 SmallVectorImpl<Function *> &Output) { 745 if (auto *Fn = Call->getCalledFunction()) { 746 Output.push_back(Fn); 747 return true; 748 } 749 750 // TODO: If the call is indirect, we might be able to enumerate all potential 751 // targets of the call and return them, rather than just failing. 752 return false; 753} 754 755static Optional<Value *> getTargetValue(Instruction *Inst) { 756 GetTargetValueVisitor V; 757 return V.visit(Inst); 758} 759 760static bool hasUsefulEdges(Instruction *Inst) { 761 bool IsNonInvokeTerminator = 762 isa<TerminatorInst>(Inst) && !isa<InvokeInst>(Inst); 763 return !isa<CmpInst>(Inst) && !isa<FenceInst>(Inst) && !IsNonInvokeTerminator; 764} 765 766static Optional<StratifiedAttr> valueToAttrIndex(Value *Val) { 767 if (isa<GlobalValue>(Val)) 768 return AttrGlobalIndex; 769 770 if (auto *Arg = dyn_cast<Argument>(Val)) 771 if (!Arg->hasNoAliasAttr()) 772 return argNumberToAttrIndex(Arg->getArgNo()); 773 return NoneType(); 774} 775 776static StratifiedAttr argNumberToAttrIndex(unsigned ArgNum) { 777 if (ArgNum > AttrMaxNumArgs) 778 return AttrAllIndex; 779 return ArgNum + AttrFirstArgIndex; 780} 781 782static EdgeType flipWeight(EdgeType Initial) { 783 switch (Initial) { 784 case EdgeType::Assign: 785 return EdgeType::Assign; 786 case EdgeType::Dereference: 787 return EdgeType::Reference; 788 case EdgeType::Reference: 789 return EdgeType::Dereference; 790 } 791 llvm_unreachable("Incomplete coverage of EdgeType enum"); 792} 793 794static void argsToEdges(CFLAliasAnalysis &Analysis, Instruction *Inst, 795 SmallVectorImpl<Edge> &Output) { 796 GetEdgesVisitor v(Analysis, Output); 797 v.visit(Inst); 798} 799 800static Level directionOfEdgeType(EdgeType Weight) { 801 switch (Weight) { 802 case EdgeType::Reference: 803 return Level::Above; 804 case EdgeType::Dereference: 805 return Level::Below; 806 case EdgeType::Assign: 807 return Level::Same; 808 } 809 llvm_unreachable("Incomplete switch coverage"); 810} 811 812// Aside: We may remove graph construction entirely, because it doesn't really 813// buy us much that we don't already have. I'd like to add interprocedural 814// analysis prior to this however, in case that somehow requires the graph 815// produced by this for efficient execution 816static void buildGraphFrom(CFLAliasAnalysis &Analysis, Function *Fn, 817 SmallVectorImpl<Value *> &ReturnedValues, 818 NodeMapT &Map, GraphT &Graph) { 819 const auto findOrInsertNode = [&Map, &Graph](Value *Val) { 820 auto Pair = Map.insert(std::make_pair(Val, GraphT::Node())); 821 auto &Iter = Pair.first; 822 if (Pair.second) { 823 auto NewNode = Graph.addNode(); 824 Iter->second = NewNode; 825 } 826 return Iter->second; 827 }; 828 829 SmallVector<Edge, 8> Edges; 830 for (auto &Bb : Fn->getBasicBlockList()) { 831 for (auto &Inst : Bb.getInstList()) { 832 // We don't want the edges of most "return" instructions, but we *do* want 833 // to know what can be returned. 834 if (auto *Ret = dyn_cast<ReturnInst>(&Inst)) 835 ReturnedValues.push_back(Ret); 836 837 if (!hasUsefulEdges(&Inst)) 838 continue; 839 840 Edges.clear(); 841 argsToEdges(Analysis, &Inst, Edges); 842 843 // In the case of an unused alloca (or similar), edges may be empty. Note 844 // that it exists so we can potentially answer NoAlias. 845 if (Edges.empty()) { 846 auto MaybeVal = getTargetValue(&Inst); 847 assert(MaybeVal.hasValue()); 848 auto *Target = *MaybeVal; 849 findOrInsertNode(Target); 850 continue; 851 } 852 853 for (const Edge &E : Edges) { 854 auto To = findOrInsertNode(E.To); 855 auto From = findOrInsertNode(E.From); 856 auto FlippedWeight = flipWeight(E.Weight); 857 auto Attrs = E.AdditionalAttrs; 858 Graph.addEdge(From, To, std::make_pair(E.Weight, Attrs), 859 std::make_pair(FlippedWeight, Attrs)); 860 } 861 } 862 } 863} 864 865static FunctionInfo buildSetsFrom(CFLAliasAnalysis &Analysis, Function *Fn) { 866 NodeMapT Map; 867 GraphT Graph; 868 SmallVector<Value *, 4> ReturnedValues; 869 870 buildGraphFrom(Analysis, Fn, ReturnedValues, Map, Graph); 871 872 DenseMap<GraphT::Node, Value *> NodeValueMap; 873 NodeValueMap.resize(Map.size()); 874 for (const auto &Pair : Map) 875 NodeValueMap.insert(std::make_pair(Pair.second, Pair.first)); 876 877 const auto findValueOrDie = [&NodeValueMap](GraphT::Node Node) { 878 auto ValIter = NodeValueMap.find(Node); 879 assert(ValIter != NodeValueMap.end()); 880 return ValIter->second; 881 }; 882 883 StratifiedSetsBuilder<Value *> Builder; 884 885 SmallVector<GraphT::Node, 16> Worklist; 886 for (auto &Pair : Map) { 887 Worklist.clear(); 888 889 auto *Value = Pair.first; 890 Builder.add(Value); 891 auto InitialNode = Pair.second; 892 Worklist.push_back(InitialNode); 893 while (!Worklist.empty()) { 894 auto Node = Worklist.pop_back_val(); 895 auto *CurValue = findValueOrDie(Node); 896 if (isa<Constant>(CurValue) && !isa<GlobalValue>(CurValue)) 897 continue; 898 899 for (const auto &EdgeTuple : Graph.edgesFor(Node)) { 900 auto Weight = std::get<0>(EdgeTuple); 901 auto Label = Weight.first; 902 auto &OtherNode = std::get<1>(EdgeTuple); 903 auto *OtherValue = findValueOrDie(OtherNode); 904 905 if (isa<Constant>(OtherValue) && !isa<GlobalValue>(OtherValue)) 906 continue; 907 908 bool Added; 909 switch (directionOfEdgeType(Label)) { 910 case Level::Above: 911 Added = Builder.addAbove(CurValue, OtherValue); 912 break; 913 case Level::Below: 914 Added = Builder.addBelow(CurValue, OtherValue); 915 break; 916 case Level::Same: 917 Added = Builder.addWith(CurValue, OtherValue); 918 break; 919 } 920 921 if (Added) { 922 auto Aliasing = Weight.second; 923 if (auto MaybeCurIndex = valueToAttrIndex(CurValue)) 924 Aliasing.set(*MaybeCurIndex); 925 if (auto MaybeOtherIndex = valueToAttrIndex(OtherValue)) 926 Aliasing.set(*MaybeOtherIndex); 927 Builder.noteAttributes(CurValue, Aliasing); 928 Builder.noteAttributes(OtherValue, Aliasing); 929 Worklist.push_back(OtherNode); 930 } 931 } 932 } 933 } 934 935 // There are times when we end up with parameters not in our graph (i.e. if 936 // it's only used as the condition of a branch). Other bits of code depend on 937 // things that were present during construction being present in the graph. 938 // So, we add all present arguments here. 939 for (auto &Arg : Fn->args()) { 940 Builder.add(&Arg); 941 } 942 943 return FunctionInfo(Builder.build(), std::move(ReturnedValues)); 944} 945 946void CFLAliasAnalysis::scan(Function *Fn) { 947 auto InsertPair = Cache.insert(std::make_pair(Fn, Optional<FunctionInfo>())); 948 (void)InsertPair; 949 assert(InsertPair.second && 950 "Trying to scan a function that has already been cached"); 951 952 FunctionInfo Info(buildSetsFrom(*this, Fn)); 953 Cache[Fn] = std::move(Info); 954 Handles.push_front(FunctionHandle(Fn, this)); 955} 956 957AliasAnalysis::AliasResult 958CFLAliasAnalysis::query(const AliasAnalysis::Location &LocA, 959 const AliasAnalysis::Location &LocB) { 960 auto *ValA = const_cast<Value *>(LocA.Ptr); 961 auto *ValB = const_cast<Value *>(LocB.Ptr); 962 963 Function *Fn = nullptr; 964 auto MaybeFnA = parentFunctionOfValue(ValA); 965 auto MaybeFnB = parentFunctionOfValue(ValB); 966 if (!MaybeFnA.hasValue() && !MaybeFnB.hasValue()) { 967 llvm_unreachable("Don't know how to extract the parent function " 968 "from values A or B"); 969 } 970 971 if (MaybeFnA.hasValue()) { 972 Fn = *MaybeFnA; 973 assert((!MaybeFnB.hasValue() || *MaybeFnB == *MaybeFnA) && 974 "Interprocedural queries not supported"); 975 } else { 976 Fn = *MaybeFnB; 977 } 978 979 assert(Fn != nullptr); 980 auto &MaybeInfo = ensureCached(Fn); 981 assert(MaybeInfo.hasValue()); 982 983 auto &Sets = MaybeInfo->Sets; 984 auto MaybeA = Sets.find(ValA); 985 if (!MaybeA.hasValue()) 986 return AliasAnalysis::MayAlias; 987 988 auto MaybeB = Sets.find(ValB); 989 if (!MaybeB.hasValue()) 990 return AliasAnalysis::MayAlias; 991 992 auto SetA = *MaybeA; 993 auto SetB = *MaybeB; 994 995 if (SetA.Index == SetB.Index) 996 return AliasAnalysis::PartialAlias; 997 998 auto AttrsA = Sets.getLink(SetA.Index).Attrs; 999 auto AttrsB = Sets.getLink(SetB.Index).Attrs; 1000 // Stratified set attributes are used as markets to signify whether a member 1001 // of a StratifiedSet (or a member of a set above the current set) has 1002 // interacted with either arguments or globals. "Interacted with" meaning 1003 // its value may be different depending on the value of an argument or 1004 // global. The thought behind this is that, because arguments and globals 1005 // may alias each other, if AttrsA and AttrsB have touched args/globals, 1006 // we must conservatively say that they alias. However, if at least one of 1007 // the sets has no values that could legally be altered by changing the value 1008 // of an argument or global, then we don't have to be as conservative. 1009 if (AttrsA.any() && AttrsB.any()) 1010 return AliasAnalysis::MayAlias; 1011 1012 return AliasAnalysis::NoAlias; 1013} 1014